Release notes

VirtuAI is updated continuously. This page lists the highlights of each month, newest first.

October 2026

Security

Stronger verification of channel webhooks. Every channel is now verified with its platform’s own mechanism: WhatsApp with Twilio’s request signature, Google Chat with Google’s signed token for your Google Cloud project, Slack with your app’s signing secret on every request, and Telegram with a secret VirtuAI generates when you register the webhook. WhatsApp and Google Chat no longer need integration keys, and integration keys now apply to voice connections. See Integration keys.
Action for Google Chat users. Google Chat requests are checked against your Google Cloud project number, GOOGLE_CHAT_PROJECT_NUMBER, which VirtuAI fills in when you save the service account JSON. Check that it’s set in Settings, and that your Chat app’s Authentication audience is Project number. See Google Chat.
Channel credentials encrypted at rest. Telegram and Slack bot tokens and secrets on agents are encrypted by VirtuAI and never returned by the API. The agent page shows them as Configured.
Tool credentials hidden from read-only users. Header values of custom tools are shown only to people who can edit tools. Everyone else sees the header names with the values masked.
Logs restricted to platform administrators. The application log viewer is no longer available to workspace admins.

New and improved

Telegram webhook managed by VirtuAI. Register Webhook now generates the webhook secret for you, and the secret field is gone from the form.
Telegram account linking. People who join a Telegram bot can link their VirtuAI account so the agent can use their connected apps. Only members of the agent’s workspace can link.

Fixes

People on Slack no longer get two answers when Slack resends an event.
Usage budgets are applied consistently on Slack.

September 2026

New and improved

Agent versioning. Edit an agent in a draft, review the changes line by line, and publish it to every channel as a numbered version. Roll back to any earlier version from the agent’s history. A workspace can require a second person to approve each publish.
Evaluations. Build datasets of test cases, or write cases by hand, and run them against an agent or its draft. Answers are scored with automatic checks or a model acting as a judge, and results are shown as charts you can compare between runs. Evaluations can also score a sample of real web chat conversations.
Conversations audit hub. Every channel’s conversations in one place, with search inside transcripts, the agent’s worklog, masked sensitive data, permission-based reveal, masked exports, and an access log of who viewed, revealed or exported what.
Single sign-on. Workspaces can sign in through their own OpenID Connect identity provider, with identity-provider groups mapped to VirtuAI roles.
Better in Gemini Enterprise. Agents called over A2A show their plan, tool calls and delegated work as progress, files an agent writes open as a Canvas, and A2A usage counts towards workspace budgets.
Thinking level. Choose Fast, Balanced or Deep for models that support reasoning, instead of each vendor’s own tiers.
Model catalog refresh. The latest Gemini, Claude and OpenAI models are available, and models retired by their providers were removed.
Choose which MCP tools are exposed. Admins can pick which of an MCP server’s tools are made available to agents.
Web chat. One searchable list for choosing the agent and model, long messages fold back into their bubble, a run’s steps appear as a single timeline, and Stop now cancels the agent’s work on the server, not only in your browser.
Google Chat. Each turn shows one card, with tool calls and delegated work grouped under the step that started them.
Analytics now covers Telegram, Slack and A2A traffic, with refreshed charts.

Fixes

Usage budgets are now enforced on Google Chat, Telegram and Slack, as they already were on web chat.
A message with an attachment that also matched a knowledge base no longer fails.
Attached files are now counted correctly towards token usage.
Restoring an agent version now shows its effect straight away.
Fixed a Gemini conversation that could stop working after a user wrote while tool results were pending.

August 2026

New and improved

A2A protocol. Expose any agent over the Agent-to-Agent protocol so other agent platforms can call it. The agent card is listed in the agent’s integration URLs, and callers authenticate with OAuth clients you create and revoke.
Slack channel. Agents can now answer in Slack.
OpenCode Go provider. Use models from DeepSeek, Qwen, GLM, Kimi, MiniMax and others through OpenCode Go, with prompt caching.
Live progress on every channel. Agents show their steps as they work, with consecutive tool calls grouped into a single card.
Long conversations. When a conversation grows past what the model can hold, earlier turns are summarized automatically and the chat tells you when it happens.

Fixes

Custom waiting messages now work in Slack and Telegram.
In Google Chat, the waiting message is replaced by the answer instead of being deleted.

July 2026

New and improved

Budgets. Set daily or monthly token and cost limits that warn or block when they are reached.
Telegram channel, with account linking so the agent knows who it is talking to.
Google Chat. Slash commands, threads, answers that stream in while the agent works, voice messages, reply context and @mentions, and a separate on/off switch for the channel.
Connected accounts for MCP servers. Users can connect their Google Workspace, GitHub, Slack or HubSpot accounts so tools act on their behalf. New ready-made Google Cloud MCP servers cover data, storage, databases and observability.
Knowledge base citations. Answers cite their sources inline, with links to the right page of a PDF, and each knowledge base has retrieval settings you can tune.
Approvals. Agents ask for permission before running shell commands.
Web chat. Token counts per message, style customization for your brand, an English and Spanish language switch for end users, and a layout that works on mobile and inside embedded frames.
Admin tools. Admins can reset a user’s password and view application logs.
New models: Claude Opus 4.8, Claude Sonnet 5 and more Gemini models on Vertex AI.
Prompt caching for supported models, with its savings reflected in cost reporting.

Fixes

Large knowledge base uploads are sent one file at a time, so big batches no longer fail.
One MCP server failing no longer stops an agent’s other tools from loading.

June 2026

New and improved

Agent Builder. Workspace admins can create and refine agents by describing what they need in a conversation.
Voice messages in web chat, with push-to-talk.
Feedback. Web chat users can rate answers with thumbs up or down, copy responses and edit their messages. Admins review ratings on the Feedback page.
Conversations show web chat and Google Chat threads, with the tools used in each.
Roles and permissions now apply consistently to every page and action.
Web chat. A rich text message box, and agent steps grouped in an expandable section.

Fixes

Web chat keeps the agent’s answer even if the connection drops before it finishes.
Google Chat posts each part of an answer as its own message.

May 2026

New and improved

Workspace web chat. A chat page for your team, with a choice of agents and a view of each agent’s tools.
Visual agent editor. A canvas for designing agents and the sub-agents they delegate to.
Vertex AI. Gemini and Claude models can run through Vertex AI.
Skills and planning. Agents can plan multi-step work, remember context, and load written procedures (skills) when a task needs them.
Secure web chat access. Web chat supports sign-in, and access can be limited to specific email domains.
Dark mode across the app.
Rich answers. Tables, code and generated files render in the chat, and a Stop button ends a response early.

April 2026

Improvements

Predefined tools now take the workspace’s time zone into account.
More reliable Google Chat replies when several messages arrive at once.

March 2026

New and improved

Analytics. Workspace analytics with drill-down and clearer charts.
Gemini 3.1 models.
Voice messages in Google Chat.
More reminder tools.

February 2026

New and improved

Predefined tools, ready to attach to agents, with configurable parameters.
Integration keys, so only trusted callers can reach an agent’s webhooks.
Invitation-only onboarding. New users join by invitation from a workspace admin.
Workspace time zone setting.
Agent form. Temperature and max output tokens settings, and an Integration URLs section with every channel URL for the agent.
Tool authentication options in the tool form.
Reminders. Agents can schedule reminders.

Fixes

Several Google Chat reliability fixes.

January 2026

New and improved

Google Chat channel.
Usage analytics for agents and workspaces.
Conversation memory is kept across restarts, so agents pick up where they left off.
Tools can receive parameters from the conversation.
Clearer agent management dialogs and confirmations.

Fixes

Agents pick up configuration changes immediately.
Fixed answers that included attachments.